Skip to content

P
PaddleHub

PaddlePaddle / PaddleHub
大约 1 年 前同步成功

通知 280
Star 12117
Fork 2091
P
PaddleHub

体验新版 GitCode，发现更多精彩内容 >>

提交 175c2005 编写于 作者: Z Zeyu Chen
浏览文件
操作

add python README and reorganize python demo

上级 1b005799
隐藏空白更改
内联 并排
Showing with 33 addition and 266 deletion
demo/mask_detection/README.md
浏览文件 @ 175c2005
# PaddleHub实现口罩佩戴检测应用
# 基于PaddleHub实现口罩佩戴检测应用
## 0 项目介绍
本文档基于飞桨本次开源的口罩佩戴识别模型, 提供了一个完整的支持视频流的WebDemo，以及高性能的Python和C++集成部署方案, 可用于不同场景下的软件集成。
## 目录
- [1. 搭建视频流场景的WebDemo](#1搭建视频流场景WebDemo)
- [2. 高性能Python部署方案](#2高性能Python部署方案)
- [3. 高性能C++部署方案](#3高性能c部署方案)
## 1. 搭建视频流场景WebDemo
![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/BB6BC87A45D146CEBA7BF237B5383835?ynotemdtimestamp=1582271320612)
### [>点击查看视频链接<](https://www.bilibili.com/video/av88962128)
##### 背景
本项目可以部署在大型场馆出入口，学校，医院，交通通道出入口，人脸识别闸机，机器人上，支持的方案有：安卓方案（如RK3399的人脸识别机，机器人），ubuntu 边缘计算，windowsPC+摄像头，识别率80%~90%，如果立项使用场景可以达到 99% （如：人脸识别机场景）。但是限于清晰度和遮挡关系，对应用场景有一些要求。
### 背景
本项目可以部署在大型场馆出入口，学校，医院，交通通道出入口，人脸识别闸机，机器人上，支持的方案有：安卓方案（如RK3399的人脸识别机，机器人），Ubuntu 边缘计算，WindowsPC+摄像头，识别率80%~90%，如果立项使用场景可以达到 99% （如：人脸识别机场景）。但是限于清晰度和遮挡关系，对应用场景有一些要求。
##### 效果分析
可以看到识别率在80~90%之前，稍小的人脸有误识别的情况，有些挡住嘴的场景也被误识别成了戴口罩，一个人带着口罩，鼻子漏出来识别成没有戴口罩，这个是合理的因为的鼻子漏出来是佩戴不规范。初步判断，这个模型应用在门口，狭长通道，人脸识别机所在位置都是可以的。
### 效果分析
可以看到识别率在80~90%之前，稍小的人脸有误识别的情况，有些挡住嘴的场景也被误识别成了戴口罩，一个人带着口罩，鼻子漏出来识别成没有戴口罩，这个是合理的因为的鼻子漏出来是佩戴不规范。这个模型应用在门口，狭长通道，人脸识别机所在位置都是可以的。
![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/7E12DBD91D1D4AB5B33C84786D519065?ynotemdtimestamp=1582271320612)![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/2BD974FB990C4C448B30B04194545054?ynotemdtimestamp=1582271320612)![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/E49E34A071F8484D948511430FAB0360?ynotemdtimestamp=1582271320612)
## 1 部署环境
### 1. 部署环境
参考： https://www.paddlepaddle.org.cn/install/quick
### 安装paddlehub
#### 安装paddlehub
`pip install paddlehub`
## 2 开发识别服务
### 加载预训练模型
### 2. 开发识别服务
#### 加载预训练模型
```python
import paddlehub as hub
module = hub.Module(name="pyramidbox_lite_mobile_mask") #口罩检测模型
......@@ -29,7 +37,7 @@ module = hub.Module(name="pyramidbox_lite_mobile_mask") #口罩检测模型
>以上语句paddlehub会自动下载口罩检测模型 "pyramidbox_lite_mobile_mask" 不需要提前下载模型
### OpenCV打开摄像头或视频文件
#### OpenCV打开摄像头或视频文件
```python
import cv2
......@@ -37,19 +45,17 @@ capture = cv2.VideoCapture(0) # 打开摄像头
# capture = cv2.VideoCapture('./2.mp4') # 打开视频文件
while(1):
ret, frame = capture.read() # frame即视频的一帧数据
if ret == False:
break
cv2.imshow('Mask Detection', frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cv2.destroyAllWindows()
```
### 口罩检测
#### 口罩佩戴检测
```python
# frame为一帧数据
......@@ -100,6 +106,7 @@ for result in results:
>原DEMO中是英文+置信度显示在框的上面，尝试改为中文，遇到字体问题，以下是解决办法
### 图片写入中文
需要事先准备ttf/otf等格式的字体文件
```python
def paint_chinese_opencv(im,chinese,position,fontsize,color_bgr):#opencv输出中文
......@@ -152,130 +159,9 @@ with open("./result/2-mask_detection.json","w") as f:
>此处可以按照自己的应用需要改为输出到mysql，Redis，kafka ，MQ 供应用消化数据
### 完整代码如下
```python
import paddlehub as hub
import cv2
from PIL import Image, ImageDraw, ImageFont
import numpy as np
import json
import os
module = hub.Module(name="pyramidbox_lite_mobile_mask")
def paint_chinese_opencv(im,chinese,position,fontsize,color_bgr):#opencv输出中文
img_PIL = Image.fromarray(cv2.cvtColor(im,cv2.COLOR_BGR2RGB))# 图像从OpenCV格式转换成PIL格式
font = ImageFont.truetype('思源黑体SC-Heavy.otf',fontsize,encoding="utf-8") # 加载字体文件
#color = (255,0,0) # 字体颜色
#position = (100,100)# 文字输出位置
color = color_bgr[::-1]
draw = ImageDraw.Draw(img_PIL)
draw.text(position,chinese,font=font,fill=color)# PIL图片上打印汉字 # 参数1：打印坐标，参数2：文本，参数3：字体颜色，参数4：字体
img = cv2.cvtColor(np.asarray(img_PIL),cv2.COLOR_RGB2BGR)# PIL图片转cv2 图片
return img
result_path = './result'
if not os.path.exists(result_path):
os.mkdir(result_path)
name = "./result/1-mask_detection.mp4"
width = 1920
height = 1080
fps = 30
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
writer = cv2.VideoWriter(name, fourcc, fps, (width, height))
maskIndex = 0
index = 0
data = []
# capture = cv2.VideoCapture(0) # 打开摄像头
capture = cv2.VideoCapture('./2.mp4') # 打开视频文件
while(1):
frameData = {}
ret, frame = capture.read() # frame即视频的一帧数据
if ret == False:
break
frame_copy = frame.copy()
完整代码可以参考`mask_detection.py`
input_dict = {"data": [frame]}
results = module.face_detection(data=input_dict)
# print(results)
maskFrameDatas = []
for result in results:
# print(result)
label = result['data']['label']
confidence_origin = result['data']['confidence']
confidence = round(confidence_origin, 2)
confidence_desc = str(confidence)
top, right, bottom, left = int(result['data']['top']), int(result['data']['right']), int(result['data']['bottom']), int(result['data']['left'])
#将当前帧保存为图片
img_name = "avatar_%d.png" % (maskIndex)
path = "./result/" + img_name
image = frame[top - 10: bottom + 10, left - 10: right + 10]
cv2.imwrite(path, image,[int(cv2.IMWRITE_PNG_COMPRESSION), 9])
maskFrameData = {}
maskFrameData['top'] = top
maskFrameData['right'] = right
maskFrameData['bottom'] = bottom
maskFrameData['left'] = left
maskFrameData['confidence'] = float(confidence_origin)
maskFrameData['label'] = label
maskFrameData['img'] = img_name
maskFrameDatas.append(maskFrameData)
maskIndex += 1
color = (0, 255, 0)
label_cn = "有口罩"
if label == 'NO MASK':
color = (0, 0, 255)
label_cn = "无口罩"
cv2.rectangle(frame_copy, (left, top), (right, bottom), color, 3)
# cv2.putText(frame, label, (left, top-10), cv2.FONT_HERSHEY_SIMPLEX, 0.8, color, 2)
frame_copy = paint_chinese_opencv(frame_copy, label_cn, (left, top-36), 24, color)
writer.write(frame_copy)
cv2.imshow('Mask Detection', frame_copy)
frameData['frame'] = index
# frameData['seconds'] = int(index/fps)
frameData['data'] = maskFrameDatas
data.append(frameData)
print(json.dumps(frameData))
index += 1
if cv2.waitKey(1) & 0xFF == ord('q'):
break
with open("./result/2-mask_detection.json","w") as f:
json.dump(data, f)
writer.release()
cv2.destroyAllWindows()
```
## 3 制作网页呈现效果
## 3.制作网页呈现效果
此DEMO是显示一个固定视频，分析导出的 json 渲染到网页里面，如需实时显示需要再次开发
### python 导出的数据
......@@ -293,11 +179,19 @@ cv2.destroyAllWindows()
![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/6329B326216A4950BF35E0CB37CDC58F?ynotemdtimestamp=1582271320612)
## 4 欢迎交流
## 2. 高性能Python部署集成方案
更多信息可以参考[文档](./python/README.md)
**百度飞桨合作伙伴：**
## 3. 高性能C++部署集成方案
更多信息可以参考[文档](./cpp/README.md)
## 欢迎交流
**百度飞桨合作伙伴：**
![image](https://note.youdao.com/yws/public/resource/b0a4695bc7d58aed3b1ff797409aee1e/DC72DE1CF51747138871BB0E3D54E20D?ynotemdtimestamp=1582271320612)
......
demo/mask_detection/mask_detection/PaddleMaskDetection.py 已删除 100644 → 0
浏览文件 @ 1b005799
# -*- coding:utf-8 -*-
import paddlehub as hub
import cv2
from PIL import Image, ImageDraw, ImageFont
import numpy as np
import json
import os
module = hub.Module(name="pyramidbox_lite_mobile_mask")
#opencv输出中文
def paint_chinese(im, chinese, position, fontsize, color_bgr):
img_PIL = Image.fromarray(cv2.cvtColor(
im, cv2.COLOR_BGR2RGB)) # 图像从OpenCV格式转换成PIL格式
font = ImageFont.truetype(
'SourceHanSansSC-Medium.otf', fontsize, encoding="utf-8")
#color = (255,0,0) # 字体颜色
#position = (100,100)# 文字输出位置
color = color_bgr[::-1]
draw = ImageDraw.Draw(img_PIL)
# PIL图片上打印汉字 # 参数1：打印坐标，参数2：文本，参数3：字体颜色，参数4：字体
draw.text(position, chinese, font=font, fill=color)
img = cv2.cvtColor(np.asarray(img_PIL), cv2.COLOR_RGB2BGR) # PIL图片转cv2 图片
return img
result_path = './result'
if not os.path.exists(result_path):
os.mkdir(result_path)
name = "./result/1-mask_detection.mp4"
width = 1920
height = 1080
fps = 30
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
writer = cv2.VideoWriter(name, fourcc, fps, (width, height))
maskIndex = 0
index = 0
data = []
capture = cv2.VideoCapture(0) # 打开摄像头
#capture = cv2.VideoCapture('./2.mp4') # 打开视频文件
while (1):
frameData = {}
ret, frame = capture.read() # frame即视频的一帧数据
if ret == False:
break
frame_copy = frame.copy()
input_dict = {"data": [frame]}
results = module.face_detection(data=input_dict)
# print(results)
maskFrameDatas = []
for result in results:
# print(result)
label = result['data']['label']
confidence_origin = result['data']['confidence']
confidence = round(confidence_origin, 2)
confidence_desc = str(confidence)
top, right, bottom, left = int(result['data']['top']), int(
result['data']['right']), int(result['data']['bottom']), int(
result['data']['left'])
#将当前帧保存为图片
img_name = "avatar_%d.png" % (maskIndex)
path = "./result/" + img_name
image = frame[top - 10:bottom + 10, left - 10:right + 10]
cv2.imwrite(path, image, [int(cv2.IMWRITE_PNG_COMPRESSION), 9])
maskFrameData = {}
maskFrameData['top'] = top
maskFrameData['right'] = right
maskFrameData['bottom'] = bottom
maskFrameData['left'] = left
maskFrameData['confidence'] = float(confidence_origin)
maskFrameData['label'] = label
maskFrameData['img'] = img_name
maskFrameDatas.append(maskFrameData)
maskIndex += 1
color = (0, 255, 0)
label_cn = "有口罩"
if label == 'NO MASK':
color = (0, 0, 255)
label_cn = "无口罩"
cv2.rectangle(frame_copy, (left, top), (right, bottom), color, 3)
# cv2.putText(frame, label, (left, top-10), cv2.FONT_HERSHEY_SIMPLEX, 0.8, color, 2)
origin_point = (left, top - 36)
frame_copy = paint_chinese(frame_copy, label_cn, origin_point, 24,
color)
writer.write(frame_copy)
cv2.imshow('Mask Detection', frame_copy)
frameData['frame'] = index
# frameData['seconds'] = int(index/fps)
frameData['data'] = maskFrameDatas
data.append(frameData)
print(json.dumps(frameData))
index += 1
if cv2.waitKey(1) & 0xFF == ord('q'):
break
with open("./result/2-mask_detection.json", "w") as f:
json.dump(data, f)
writer.release()
cv2.destroyAllWindows()
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或